Quality-Adjusted Prices for the American Motor Automobile Industry 1906-1940:

Considering Cars as New Goods

Daniel M.G. Raff, The Wharton School and NBER, and Manuel Trajtenberg, Tel-Aviv University and NBER

Please see the longer version of this paper (available on request from raff@wmgtfac.wharton.upenn.edu) for a lengthy list of people and institutions to whom we owe thanks, the usual disclaimer, and much more extensive text, tables, graphs, footnotes, and scholarly apparatus.

1. Introduction

The empirical literature on new goods has long shown an interest in the automobile. The hedonic approach was introduced to the profession in Court's 1939 attempt to measure the evolution of automobile prices on a quality-adjusted basis. Griliches 1961, Triplett 1969, Ohta and Griliches 1974, and Gordon 1990 (Ch. 8), all leading references, continued the study of the industry. Yet each of these, only Court excepted, is focused on developments that took place in the years after the Great Depression, a period when the automobile as an innovation was clearly mature. Trajtenberg 1990, a recent study of CT scanners, suggests that the largest contributions of new goods to welfare changes may well come much earlier on. The industry's annual model changes pose the price index question perfectly well in the post-war period. But the most salient new goods questions necessarily take us further back in time.

Straightforward facts about the history of automobile manufacturing in America support this view. In the first three decades of this century, the industry went from scarcely existing, insofar as Census of Manufactures enumerators were concerned, to being in terms of the value of products the largest industry in the economy. Over the period we study in this paper, the most casual observer can recognize how much the product changed. Manufacturing methods evolved equally dramatically. So too did market prices. In 1906, for example, there were no new automobiles for sale at a price equal to or below the GNP per capita at the time of $336. In fact, the average price in our database for that year is nearly ten times that. By 1940, where our data end, a household with a year's GNP per capita to spend ($754) had a choice of 59 different models, and the average price of cars on the market that year was only about twice that sum.

The industry saw tremendous changes over this period as well. Indeed, contrasted to the tight oligopoly and dull performance of the post-World War II decades, the vibrancy of these early years is almost shocking. There was an early and well-organized attempt at cartelization that failed. Entry eventually proceeded at a breakneck pace. Attracted by the palpably vast opportunities, hundreds of new firms burst onto the scene every year, the total running to well in excess of a thousand. More than 10,000 distinct models were on offer at one time or another. Intense competition in price and quality persistently pushed price-performance ratios to new lows.

The consequences were far-reaching. One advertising slogan early in the period ran "One day, one dollar; one year, one Ford". In the very beginning, automobiles were strictly playthings of the rich. But well before 1940, cars were routinely purchased by ordinary working households. The consequences of this for American economic life were themselves pervasive and profound. At the turn of the century, even private urban transportation was often powered by horses. Roads were often dusty when dry and all but impassable when wet. But by 1940, the internal combustion engine ruled the way. Road construction techniques were recognizably modern. All-weather paved roads existed all over the nation. Bedroom suburbs and even places of work and trade were located in areas where trains and trolleys did not run. The automobile was a new good with important consequences.

There is a vast literature on the industry's history. However, quite surprisingly, it contains no systematic quantitative analysis of the period in which most of the technical change happened. Price indices would be a useful start. We proceed in steps. The first is simply to complement the existing hedonic literature by pushing the span of its automobile quality-adjusted price calculations backwards to 1906, thus closing in on the birth of the industry and product. These price indices can be used for a variety of purposes. We propose a crude decomposition of the price change into product- and process-innovation components, identifying quality-constant price change with manufacturing economies and quality change with design improvement. We also use our indices for comparisons to hedonic price indices for other industries at a comparably early stage of the product life cycle and for comparisons to hedonic price indices for this industry in the later periods previously studied. In particular, we couple our results to those of Gordon's analogous exercise for the post-World War II period to consider the industry's history in the long view. Finally, we assess several possible sources of bias in our results.

The paper proceeds in six main parts. Section 2 (omitted from this text) is a technical introduction to the product. Section 3 (largely intact) discusses the data. Section 4 (largely omitted) gives preliminaries to the hedonic analysis and discusses the regressions. Section 5 (largely intact) presents the main results in terms of quality-adjusted price indices and puts them in the wider context. Section 6 (entirely omitted) considers the seriousness of two potential sources of bias in the index numbers (and finds them provisionally untroubling). Section 7 (entirely omitted) concludes. There is also an omitted appendix.

2. Cars: A Technical Overview


3. The Data

Computing quality-adjusted prices indices, even using as undemanding a method as the hedonic, requires large amounts of very detailed data. One needs prices and detailed attribute information for virtually all the different models marketed in each period. Studies such as this thus rest firmly on the breadth of their data.

The primary source of most information about the identities and systems of individual models that covers any wide range of models is the specification tables published in the contemporary trade press at the time of the annual New York Auto Show. The trade journals vary in the attributes they report. The attributes reported in each source also change slowly over time. The information given about some attributes is not as revealing as it might be. The tables are nonetheless very detailed and an extremely rich data source.

Each mechanically distinct variant identified in the tables could usually be purchased with any of several different bodies. We call these pairings body models and use them as our unit of observation. We were constrained (by time and finances) to enter body-model data only for alternate years and to go back no further than 1906. Table 3.1 (omitted from this text) gives some basic descriptive statistics. We have a total of over 11,000 observations (i.e. of body-models offered). The number rises sharply in the earliest years, more through entry than through model proliferation. It peaks in 1910 at 1006. There is a second surge after WWI and a third at the end of the twenties, after which the number declines considerably. There was a pronounced decline in the number of manufacturers over the period and substantial model proliferation in the thirties.


After some research, we concluded that the attributes reported by the periodicals Automotive Industries and Motor together generally spanned the information available. We thus drew the data on attributes and prices from these periodicals. Coverage was then compared against the listings in Kimes and Clark 1985, apparently the most authoritative hobbyist source. Spot checks with other researchers and comparisons with industry histories and other such investigations covering this period, published and otherwise, have revealed no important or systemically unutilized information. It is important to note that our data represent only firms operating above a certain minimal economic threshold, namely they were large enough to make advertising at the major annual trade show attractive. We may have thus left out experimentalists and bespoke manufacturers so aloof from commerce that they left customers to find their own way to the factory. We have surely left out some hopeful entrepreneurs who had and possibly even announced bold plans but never in fact made any cars. But we have found no evidence that we have left out any products that were actually easy to buy, however, and this is the breadth of data that the hedonic method requires.

4. Hedonic analysis: preliminaries


4.2 Selection of attributes

One fundamental difficulty has beset all hedonic car studies from Court onwards. It is that of identifying a set of attributes that can be taken to be the most important performance attributes of cars and that can be measured in a consistent fashion over time. Only if quality in this sense is quite tightly controlled for can we begin to regard as reliable quality-adjusted price indices based on hedonic regressions.

Any quality adjustment method requires regressors that would in principle go directly into a consumer's utility function. "Reliability," "smoothness of ride," "safety," "comfort," etc. are presumably the sort of attributes in question. But these are extremely hard to quantify in an objective or even consistent manner. Engineering (i.e. technical) attributes are much easier to measure; but they are certainly further removed from the quality dimensions perceived by consumers.

The difficulty in identifying structural relationships between engineering attributes and utility stems from the fact, sketched in Section 2 above, that for all their pervasiveness and ease of operation, cars are extremely complex machines. Their overall performance depends in a complicated way upon the performance of each of the systems, upon trade-offs made between them, and upon the extent to which their design is well integrated. All this makes it a formidable challenge to devise variables that would even proxy the performance of individual model designs in an unambiguous and parsimonious way. We have made some progress in that respect in this study by including (apparently for the first time) actual measures for many of those systems (brakes, clutch, drive mechanism, etc.). Whether our selection of systems and variable definitions are the most appropriate or effective only further investigation will reveal.

In the end, we decided to include three categories of attributes in the hedonic regressions: measures of vehicle size, engine power, and the technology of five major engineering systems. Size and power have been used in virtually all auto hedonics studies. They are very closely associated with price; and casual empiricism suggests that consumers do care about them. For systems, we initially attempted to cover all the major ones identified in Section 2. In particular pairs of years, however, we often had to make significant compromises in the face of data limitations of various sorts.

For size we use wheel base, measured in inches. For power we have available for most years two alternative measures: rated horsepower and displacement. We opt for the latter whenever it is available because it captures more information (i.e. stroke, bore, and number of cylinders). The five systems we chose are the rear axle, clutch, brakes, drive type, and suspension. The dummy variables are defined in Table 4.1 (omitted here) with their names as they appear in the hedonic regression results (also omitted).

Each of these systems underwent dramatic changes over the period studied. Technical innovations, changes in demand, and the shifting interactions with related systems made particular designs emerge and diffuse, only to be superseded later by others. The methods of this project required us to trace and grasp the evolution of system designs over time, both in order to define the categories that eventually appeared as dummy variables in the hedonic regressions and to form priors as to the likely sign of their coefficients. In addition, we believe that the time paths followed by competing designs are of significant interest in themselves. They show vividly the contest between alternative systems and the speed of diffusion of those that emerged as dominant. We present in the Appendix (here omitted) a technical and graphical description of the evolution of the main systems.

If one of the types should become a virtual standard (i.e. if its share among the competing models approaches 100%), then it approaches collinearity with the regressions' constant terms. The system can no longer be included in the regression. That is the case for spring type from 1928 on, for example: the half-elliptical type had been adopted by then in over 95% of all cars marketed. In other cases, though, one type became dominant but then differentiates as subvariants appear. In this case, the system can be still included: it merely requires a different dummy variable. For example, by 1928 the dominant clutch type was plate, but for a few years afterwards the market split between single plate and double plate. In the case of the drive type, the spiral bevel acquired absolute dominance by 1922; but from 1926 on it had to compete against the hypoid type. By 1940 the latter was present in 80% of all models.


5. Quality-Adjusted Price Indices


5.1 Simple quality-adjusted price indices

On the basis of the estimated hedonic coefficient, denoted hereafter by a, we compute a quality-adjusted percentage price change as follows.

%ÆQAPrice = 1 - exp([[alpha]])

QA stands here for quality-adjusted, %Æ for percentage change. We calculate %ÆQAPrice both for [[alpha]]'s estimated on the basis of current prices and for [[alpha]]'s estimated on the basis of CPI-deflated prices. We then construct corresponding quality-adjusted price indices with the results shown in Table 5.1.

The main findings are as follows. First, quality-adjusted prices (based on CPI-deflated prices) fell at an average rate of slightly more than 5% per year from 1906 to 1940, thus halving every 13 years. This is by absolute standards quite a substantial pace. In constant-1993 dollar terms, it means that the average price of a car of constant quality was $52,600 in 1906 and that this fell to just $8,100 by 1940. To put this in perspective, if the industry would have continued to innovate at the same rate from 1944 to the present (1994), a car would cost these days just $582 on a quality-adjusted basis.

Secondly, as is to be expected, the rate of change of quality-adjusted prices using CPI-deflated prices was generally larger in absolute value than using current prices. The exception is periods of marked deflation, during which auto prices - like the prices of many durables - dropped more slowly than the CPI. Thirdly, we ran different variants of the hedonic regressions and constructed the corresponding indices in order to ascertain the role played by the inclusion of the variables representing the five engineering systems. The results (not shown in the tables) indicate that their inclusion does make a difference, but for the most part, it is a small one - in the range of one-half to one and a half percentage points per year in the computation of %ÆQAPrice.

5.2 Process- versus product innovation

We next compute a rate of quality change, defined as a residual:

%ÆQuality = %ÆPrice - %ÆQAPrice

If the attributes of cars remain constant, %ÆPrice is exactly equal to %ÆQAPrice, and %ÆQuality must take the value of zero. Suppose, on the other hand, that cars improve. Then %ÆQAPrice is strictly less than %ÆPrice. We might call the difference--that is, %ÆQuality - pure quality change. If there is technical advance then this difference would be positive. (In this case %ÆQAPrice would be negative, since it refers to the quality-adjusted price decline). Notice that %ÆQuality can take negative values if quality-adjusted prices drop less or rise more than unadjusted prices. That would be the case, for example, if prices did not change but some cars displayed less of some attributes that were positively valued (or, more precisely, that show a positive coefficient in the hedonic regression)

The series is displayed in Table 5.2. The 5% average annual decline of quality-adjusted prices can be decomposed as follows. Prices by themselves (CPI-deflated) dropped at a rate of 3% per year. The residual "quality" therefore increased at a rate of 2%. If we identify quality-constant price change with manufacturing economies and quality change as we have defined it with design improvements, then these numbers suggest that 60% of the decline in quality-adjusted prices was due to process innovation and only 40% to product innovation or quality change per se.

This partition of the overall quality-adjusted price decline into a product innovation and a process innovation component should be regarded cautiously. Many modern manufacturing economies, for example, come from simplifying designs, and a reliable decomposition would therefore have to study specific innovations (see for example Whitney 1988). And prices can certainly fall for a variety of reasons, among them increased competition and lower input prices. But the identification with process innovation seems plausible because of the dramatic economies offered by the development and diffusion of mass production methods. There can be no doubt that the set of techniques grouped under the umbrella term of mass production constituted one of the most important innovations in manufacturing methods of all time and had tremendous consequences in terms of unit costs, scale, and production capabilities. The drop from, say, the $2,000-$3,000 cars of the early 1910's to the sub-$500 Ford Model T would have never been possible with the craft-like production and assembly methods that prevailed then.

It remains to be established, however, precisely how much of the industry's overall price drop could be attributed to the diffusion of mass production and what exactly the causal link was. Casual evidence suggests that the relationship was very non-linear, perhaps because of the interplay between innovation and competition. Recall that prices dropped a great deal in the immediate aftermath of Ford's introduction of mass production. Recall also that this was a period in which Ford was the only producer to operate in this fashion. We speculated above that the generalized drop was due to competitive pressures brought about by Ford's drastic price reductions. That the downward trend in prices continued along with the diffusion of mass production is certainly consistent with this explanation;, but it is not clear how closely synchronized the two processes were. It would also be interesting to see whether the steep and sustained drop in prices experienced by the automobile industry over more than three decades is typical of new industries along their trajectory towards maturity or was unique.

5.3 Quality-adjusted price changes over sub-periods

As Table 5.3 reveals, one can clearly distinguish four periods in terms of %_QAPrice and %_Quality. Note that the partition is not exactly the same for the two measures. Most of the innovation appears to have occurred very early on (i.e. 1906 through either 1914 or 1918, depending on which series one uses). Moreover, the highest rates of quality change occurred at the very beginning (1906-14). This is undoubtedly the portion of our period in which the greatest proportion of entrepreneurs were engineers or mechanics by training, knowledge spillovers were all pervasive, and design bureaucracies were shallowest. Whatever the mechanisms may have been, the pattern lends further support to the conjecture that it is indeed in the course of the emergence of a new industry that the largest strides in product innovation are made. An important implication of this is that if one misses out those early stages in computing quality-adjusted price indices, one is bound to grossly underestimate the welfare effects of product innovation.

In order to gain some perspective on the observed rate of innovation in cars during the initial period, it is worth comparing it to what might be regarded as the parallel period for personal computers, namely 1982-88. As reported in Berndt and Griliches 1993, the average rate of quality-adjusted price decline in that industry and period was somewhere between -0.20 and -0.30% per year (depending on the sort of estimate used). For cars, our results show a figure of about half that size (-0.11% per year for 1906-18, -0.14 per year for 1906-14). This is quite remarkable considering that the case of personal computers is widely regarded as extreme in its rate of real price decline. The decline for PC's derived primarily from a long and steady series of dramatic improvements in integrated circuit - in particular, microprocessor - design and manufacturing capabilities. No major automobile component experienced such sustained dramatic price/performance declines. Yet the entire choice spectrum of cars displayed 11 to 14% yearly rates of quality adjusted price drops for roughly a decade!


5.4 A longer horizon

It is natural to want to place the main findings of this Section in the context of a more extended history of the industry. The obvious way to do this is to link our series to the recent series of Gordon running from just after the war through the early 1980s. Since that series also derives from unweighted regressions, it is in fact appropriate to link the two directly. The linking can be accomplished using numbers relating 1937 and 1950 cross-sections from Griliches 1961. Table 5.4 gives the combined series. Figure 5.1 (omitted here but to be displayed during my five minutes in San Francisco) illustrates.

It would be in the spirit of the literature to give a detailed interpretation to this Figure. But radical changes in the general price level over this extended period suggest deflating by the CPI first. This yields the series illustrated in Figure 5.2 (omitted here but, again, to be displayed in the presentation). The explosion at the end of the series in Figure 5.1 - proportionately roughly as large as the declines of the early years - is revealed to be for practical purposes entirely inflation. The overwhelming bulk of the quality-adjusted price decline in this industry came in a tremendous burst before the '20s. By the time the Depression was over, so was 90% of the story. Computations of growth rates averaged out over very long intervals can indeed miss the most salient details. The action here came when the good was new.

6. Potential Biases


7. Conclusion




Rate of change Index

using: using:

Current Constant Current Constant

prices prices prices prices

1908 -0.30 -0.30 70.0 70.0

1910 -0.23 -0.25 54.0 52.4

1912 -0.09 -0.12 49.3 46.0

1914 -0.12 -0.16 43.3 38.8

1916 -0.24 -0.30 33.1 27.4

1918 0.16 -0.16 38.4 23.1

1920 0.36 0.03 53.4 23.8

1922 -0.09 0.09 47.9 25.9

1924 -0.14 -0.16 41.2 21.9

1926 -0.10 -0.13 37.3 19.0

1928 -0.07 -0.12 34.8 16.7

1930 -0.04 -0.10 33.4 15.0

1932 -0.18 0.00 27.4 14.9

1934 -0.08 -0.06 25.3 14.0

1936 0.01 -0.02 25.5 13.8

1938 0.19 0.16 30.2 16.0

1940 -0.05 -0.04 28.8 15.4

Ann. avg. 1906-1940 -0.03 -0.05

"Constant" prices are CPI-deflated, 1993=100



(IN CONSTANT 1993 $)

Mean Mean %_ %_ %_

Year Price QAPrice Price QAPrice Quality

1906 52 640 52 640

1908 46 640 36 848 -0.11 -0.30 0.19

1910 39 860 27 583 -0.15 -0.25 0.10

1912 41 400 24 214 0.04 -0.12 0.16

1914 44 242 20 424 0.07 -0.16 0.23

1916 29 483 14 423 -0.33 -0.30 -0.03

1918 24 875 12 160 -0.16 -0.16 0.00

1920 24 566 12 528 -0.01 0.03 -0.04

1922 27 146 13 634 0.11 0.09 0.02

1924 22 732 11 528 -0.16 -0.16 0.00

1926 22 082 10 002 -0.03 -0.13 0.10

1928 21 241 8 791 -0.04 -0.12 0.08

1930 20 702 7 896 -0.03 -0.10 0.07

1932 25 803 7 843 0.25 0.00 0.25

1934 23 236 7 370 -0.10 -0.06 -0.04

1936 17 842 7 264 -0.23 -0.02 -0.21

1938 19 036 8 422 0.07 0.16 -0.09

1940 16 565 8 107 -0.13 -0.04 -0.09

Avg. 1906-1940 -0.06 -0.10 0.04

QA: Quality-adjustment from the hedonic regressions



(i) %Æ QAPrice

1906-1918: -0.22

1918-1922: 0.06

1922-1930: -0.13

1930-1940: 0.01

1906-1940: -0.10

(ii) %Æ Quality

1906-1914: 0.17

1914-1924: -0.01

1924-1932: 0.12

1930-1940: -0.11

1906-1940: 0.04



1906 100.0

1908 70.0

1910 54.0

1912 49.3

1914 43.3

1916 33.1

1918 38.4

1920 53.4

1922 47.9

1924 41.2

1926 37.3

1928 34.8

1930 33.4

1932 27.4

1934 25.3

1936 25.5

1938 30.2

1940 28.8

1947 34.7

1948 39.9

1949 46.9

1950 45.0

1951 48.8

1952 49.7

1953 49.7

1954 48.3

1955 50.8

1956 49.7

1957 50.3

1958 49.7

1959 50.8

1960 50.3

1961 50.8

1962 52.8

1963 51.8

1964 51.3

1965 50.3

1966 50.8

1967 51.3

1968 53.4

1969 52.3

1970 53.9

1971 57.8

1972 55.6

1973 54.5

1974 58.4

1975 68.5

1975 72.0

1977 74.3

1978 85.4

1979 88.9

1980 99.2

1981 124.9

1982 135.3

1983 140.8

The Coefficient on the variable D in Table 4 of Griliches 1961 was used to splice the third column of our Table 4.4 and Column 6 of Table 8.8 in Gordon 1990.